增强技能系统与记忆机制，完善文档说明

- 支持外部技能插件加载，技能与命令工具集成
- 工作记忆文件改为 JSON 格式，优化结构化存储
- 补充记忆注入机制及各类 memory key 说明
- MainAgent 工具调用追踪范围限定为当前轮次
- 对话上下文自动注入当前时间，便于记忆检索
- 配置示例新增 recaller agent，文档同步 .NET 10 正式版
- 行为准则新增技能自省，禁止虚构未授权操作
- 全面更新中英文文档，提升可扩展性与规范性

Author: xcssa

CLAUDE.md

@@ -1,4 +1,4 @@
-# CLAUDE.md — Sharpclaw Project Instructions
+# CLAUDE.md
 
 This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
 
@@ -91,7 +91,7 @@ Sharpclaw implements a sophisticated three-layer memory pipeline:
 
 | Layer | File | Purpose | Written By |
 |-------|------|---------|------------|
-| **Working Memory** | `working_memory.md` | Current conversation snapshot | MainAgent (each turn) |
+| **Working Memory** | `working_memory.json` | Current conversation snapshot (JSON) | MainAgent (each turn) |
 | **Recent Memory** | `recent_memory.md` | Detailed summaries (append-only) | ConversationArchiver (Summarizer) |
 | **Primary Memory** | `primary_memory.md` | Consolidated core facts | ConversationArchiver (Consolidator) |
 | **Vector Store** | `memories.db` | Semantic embeddings + metadata | VectorMemoryStore |
@@ -102,6 +102,15 @@ Sharpclaw implements a sophisticated three-layer memory pipeline:
 2. When context window overflows → Summarizer generates detailed summary → appends to recent memory
 3. When recent memory > 30k chars → Consolidator extracts core info → overwrites primary memory
 
+**Memory Injection Mechanism:**
+- Injected messages are identified via `AdditionalProperties` dictionary keys:
+  - `AutoMemoryKey` (`__auto_memories__`) — Vector search results
+  - `AutoRecentMemoryKey` (`__auto_recent_memory__`) — Recent memory content
+  - `AutoPrimaryMemoryKey` (`__auto_primary_memory__`) — Primary memory content
+  - `AutoWorkingMemoryKey` (`__auto_working_memory__`) — Working memory snapshot
+- Memory messages are replaced/updated each turn rather than accumulated
+- Turn-based message tracking via `turnStartIndex` separates current turn from full history
+
 ### IChatIO Abstraction
 
 The AI engine is decoupled from frontend through `Abstractions/IChatIO.cs`:
@@ -111,6 +120,14 @@ The AI engine is decoupled from frontend through `Abstractions/IChatIO.cs`:
 
 All frontends share the same `MainAgent` logic.
 
+### Skills System
+
+External skills are loaded from `~/.sharpclaw/skills/` via `AgentSkillsDotNet`:
+- `AgentBootstrap` scans the skills directory at startup and converts discovered skills into `AITool[]`
+- Skills are merged with built-in command tools into a unified `CommandSkills` array
+- All tools (built-in commands + external skills) are passed to `MainAgent` as a single collection
+- The skills directory is auto-created if it doesn't exist
+
 ---
 
 ## Configuration
@@ -144,6 +161,9 @@ Configuration stored in `~/.sharpclaw/config.json` (version 8):
 ### Sharpclaw
 - `Microsoft.Agents.AI.OpenAI` / `Microsoft.Extensions.AI.OpenAI` — AI agent framework
 - `Anthropic` / `GeminiDotnet.Extensions.AI` — Multi-provider LLM support
+- `AgentSkillsDotNet` — External skills loading and tool generation
+- `ToonSharp` — Lua scripting support
+- `Cronos` — Cron expression scheduling
 - `Terminal.Gui` v2 (develop) — TUI framework
 - `Luolan.QQBot` — QQ Bot SDK
 - `Microsoft.Data.Sqlite` — SQLite for vector store
@@ -202,15 +222,16 @@ sharpclaw/
 - All sub-agents use **tool calling** (not text parsing) for structured output
 - Sub-agents access memory files via standard file command tools with prompt-level write restrictions
 - Injected messages use `AdditionalProperties` dictionary keys for identification
-- Session persistence via `working_memory.md`, `memories.db`, tiered memory files
+- Session persistence via `working_memory.json`, `memories.db`, tiered memory files
 
 ### Agent Behavior (From MainAgent System Prompt)
 1. **Context-Aware Execution**: Use `CommandDir`/`FindFiles` when unfamiliar with project; skip if memory sufficient
 2. **Read Before Write**: Always `CommandCat` before `CommandEditText` to get exact line numbers
 3. **Evaluate Impact**: Use `SearchInFiles` before modifying public APIs
-4. **Goal Decomposition**: Break complex tasks into subtasks, execute continuously
-5. **Auto-Recovery**: On failure, analyze stderr, retry 2-3 times before reporting
-6. **Strict Verification**: Check Git Diff after edits, fix immediately if wrong
+4. **Skill Introspection**: Check available tools/skills before claiming capabilities; never fabricate unauthorized operations
+5. **Goal Decomposition**: Break complex tasks into subtasks, execute continuously
+6. **Auto-Recovery**: On failure, analyze stderr, retry 2-3 times before reporting
+7. **Strict Verification**: Check Git Diff after edits, fix immediately if wrong
 
 ---
 

README.md

@@ -29,6 +29,12 @@ By leveraging the `Microsoft.Extensions.AI` abstraction layer, Sharpclaw seamles
 * **Web (WebSocket):** A lightweight ASP.NET Core web server with a modern UI (Tokyo Night theme) and real-time streaming.
 * **QQ Bot:** Native integration with QQ channels, groups, and private messages.
 
+### 🔌 Extensible Skills System
+
+* **External Skills:** Load custom skills from `~/.sharpclaw/skills/` via `AgentSkillsDotNet`, seamlessly merged with built-in commands as a unified tool collection.
+* **Lua Scripting:** Embedded scripting support via `ToonSharp` for user-defined automation.
+* **Scheduled Tasks:** Cron expression support via `Cronos` for timed operations.
+
 ### 🔒 Secure Configuration
 
 * Cross-platform secure credential storage (Windows Credential Manager, macOS Keychain, Linux libsecret) using AES-256-CBC encryption for API keys.
@@ -41,7 +47,7 @@ By leveraging the `Microsoft.Extensions.AI` abstraction layer, Sharpclaw seamles
 
 ### Prerequisites
 
-* [.NET 10.0 SDK (Preview)](https://dotnet.microsoft.com/)
+* [.NET 10.0 SDK](https://dotnet.microsoft.com/)
 * Git (for cloning submodules)
 
 ### Build and Run
@@ -107,15 +113,13 @@ dotnet run --project sharpclaw config
 │  ├── VectorMemoryStore — Sharc + SQLite vector search       │
 │  └── InMemoryMemoryStore — Keyword-based fallback           │
 ├─────────────────────────────────────────────────────────────┤
-│  Command System (Commands/)                                  │
-│  ├── FileCommands — File operations (cat, edit, find, etc.) │
-│  ├── ProcessCommands — Bash/PowerShell execution            │
-│  ├── HttpCommands — HTTP requests                           │
-│  ├── TaskCommands — Background task management              │
-│  └── SystemCommands — System info, exit                     │
+│  Skills & Commands (Commands/, ~/.sharpclaw/skills/)         │
+│  ├── Built-in — File, Process, HTTP, Task, System commands  │
+│  ├── External Skills — AgentSkillsDotNet plugin loading     │
+│  └── Memory Tools — SearchMemory, GetRecentMemories         │
 ├─────────────────────────────────────────────────────────────┤
 │  Core Infrastructure (Core/)                                 │
-│  ├── AgentBootstrap — Shared initialization                 │
+│  ├── AgentBootstrap — Shared initialization + skill loading │
 │  ├── SharpclawConfig — Configuration with encryption        │
 │  ├── ClientFactory — LLM client creation                    │
 │  ├── DataProtector/KeyStore — AES-256-CBC encryption        │
@@ -129,7 +133,7 @@ Sharpclaw implements a sophisticated three-layer memory pipeline:
 
 | Layer | File | Purpose |
 |-------|------|---------|
-| **Working Memory** | `working_memory.md` | Current conversation snapshot |
+| **Working Memory** | `working_memory.json` | Current conversation snapshot |
 | **Recent Memory** | `recent_memory.md` | Detailed summaries (append-only) |
 | **Primary Memory** | `primary_memory.md` | Consolidated core facts |
 | **Vector Store** | `memories.db` | Semantic embeddings + metadata |
@@ -195,6 +199,7 @@ Configuration is stored in `~/.sharpclaw/config.json` (version 8):
   },
   "agents": {
     "main": { "enabled": true },
+    "recaller": { "enabled": true },
     "saver": { "enabled": true },
     "summarizer": { "enabled": true }
   },

README_CN.md

@@ -29,6 +29,12 @@ Sharpclaw 是一个基于 **.NET 10** 开发的**自主型 AI 智能体框架**
 * **Web 模式 (WebSocket):** 轻量级 ASP.NET Core 服务，内置现代化网页终端（Tokyo Night 暗色主题），支持流式打字机输出。
 * **QQ Bot 模式:** 原生接入 QQ 机器人体系，支持频道、群聊及私聊环境。
 
+### 🔌 可扩展技能系统
+
+* **外部技能加载:** 从 `~/.sharpclaw/skills/` 目录通过 `AgentSkillsDotNet` 加载自定义技能，与内置命令无缝合并为统一工具集。
+* **Lua 脚本支持:** 通过 `ToonSharp` 嵌入式 Lua 引擎实现用户自定义自动化脚本。
+* **定时任务:** 通过 `Cronos` 支持 Cron 表达式调度定时操作。
+
 ### 🔒 高安全性配置
 
 * 跨平台凭据安全存储：支持 Windows 凭据管理器、macOS 钥匙串和 Linux libsecret，API Key 等敏感信息默认通过 AES-256-CBC 加密存储。
@@ -41,7 +47,7 @@ Sharpclaw 是一个基于 **.NET 10** 开发的**自主型 AI 智能体框架**
 
 ### 环境要求
 
-* [.NET 10.0 SDK (Preview)](https://dotnet.microsoft.com/)
+* [.NET 10.0 SDK](https://dotnet.microsoft.com/)
 * Git (用于拉取子模块)
 
 ### 编译与运行
@@ -107,15 +113,13 @@ dotnet run --project sharpclaw config
 │  ├── VectorMemoryStore — Sharc + SQLite 向量搜索            │
 │  └── InMemoryMemoryStore — 关键字匹配降级方案               │
 ├─────────────────────────────────────────────────────────────┤
-│  命令系统 (Commands/)                                        │
-│  ├── FileCommands — 文件操作（cat、edit、find 等）          │
-│  ├── ProcessCommands — Bash/PowerShell 执行                 │
-│  ├── HttpCommands — HTTP 请求                               │
-│  ├── TaskCommands — 后台任务管理                            │
-│  └── SystemCommands — 系统信息、退出                        │
+│  技能与命令系统 (Commands/, ~/.sharpclaw/skills/)             │
+│  ├── 内置命令 — 文件、进程、HTTP、任务、系统                │
+│  ├── 外部技能 — AgentSkillsDotNet 插件加载                  │
+│  └── 记忆工具 — SearchMemory、GetRecentMemories             │
 ├─────────────────────────────────────────────────────────────┤
 │  核心基础设施 (Core/)                                        │
-│  ├── AgentBootstrap — 共享初始化逻辑                        │
+│  ├── AgentBootstrap — 共享初始化 + 技能加载                 │
 │  ├── SharpclawConfig — 带加密的配置管理                     │
 │  ├── ClientFactory — LLM 客户端创建                         │
 │  ├── DataProtector/KeyStore — AES-256-CBC 加密              │
@@ -129,7 +133,7 @@ Sharpclaw 实现了复杂的三层记忆管线：
 
 | 层级 | 文件 | 用途 |
 |------|------|------|
-| **工作记忆** | `working_memory.md` | 当前会话快照 |
+| **工作记忆** | `working_memory.json` | 当前会话快照 |
 | **近期记忆** | `recent_memory.md` | 详细摘要（追加模式） |
 | **核心记忆** | `primary_memory.md` | 整合后的核心事实 |
 | **向量库** | `memories.db` | 语义嵌入 + 元数据 |
@@ -195,6 +199,7 @@ sharpclaw/
   },
   "agents": {
     "main": { "enabled": true },
+    "recaller": { "enabled": true },
     "saver": { "enabled": true },
     "summarizer": { "enabled": true }
   },

sharpclaw/Agents/MainAgent.cs

@@ -222,6 +222,7 @@ void Append(string type, string text)
         // 流式输出
         _reducer.UserInput = input;
         _reducer.WorkingMemoryBuffer.Add(new ChatMessage(ChatRole.User, input));
+        var turnStartIndex = _reducer.WorkingMemoryBuffer.Count; // 当前轮次流式消息的起始索引
         AIContent? lastContent = null;
         AppLogger.SetStatus("AI 思考中...");
         _chatIO.BeginAiResponse();
@@ -294,14 +295,16 @@ void Append(string type, string text)
         if (message.Contents.Count > 0)
             _reducer.WorkingMemoryBuffer.Add(message);
 
-        // 检查所有未返回结果的工具调用，补充取消标记
-        var allCallIds = _reducer.WorkingMemoryBuffer
+        // 检查当前轮次中未返回结果的工具调用，补充取消标记
+        var turnMessages = _reducer.WorkingMemoryBuffer.Skip(turnStartIndex);
+
+        var allCallIds = turnMessages
             .Where(m => m.Role == ChatRole.Assistant)
             .SelectMany(m => m.Contents.OfType<FunctionCallContent>())
             .Select(c => c.CallId)
             .ToHashSet();
 
-        var answeredCallIds = _reducer.WorkingMemoryBuffer
+        var answeredCallIds = turnMessages
             .Where(m => m.Role == ChatRole.Tool)
             .SelectMany(m => m.Contents.OfType<FunctionResultContent>())
             .Select(r => r.CallId)
@@ -310,7 +313,7 @@ void Append(string type, string text)
         allCallIds.ExceptWith(answeredCallIds);
         if (allCallIds.Count > 0)
         {
-            var unmatchedCalls = _reducer.WorkingMemoryBuffer
+            var unmatchedCalls = turnMessages
                 .Where(m => m.Role == ChatRole.Assistant)
                 .SelectMany(m => m.Contents.OfType<FunctionCallContent>())
                 .Where(c => allCallIds.Contains(c.CallId))

sharpclaw/Chat/MemoryPipelineChatReducer.cs

@@ -95,6 +95,13 @@ public async Task<IEnumerable<ChatMessage>> ReduceAsync(
         if (_systemPrompt is not null && systemMessages.Count == 0)
             systemMessages.Add(new ChatMessage(ChatRole.System, _systemPrompt));
 
+        // 注入当前时间
+        systemMessages.Add(new ChatMessage(ChatRole.System,
+            $"[当前时间] {DateTimeOffset.Now:yyyy-MM-dd HH:mm:ss zzz}")
+        {
+            AdditionalProperties = new() { [AutoMemoryKey] = true }
+        });
+
         // ── 2. 保存工作记忆（从 MainAgent 流式累积的 WorkingMemoryBuffer）──
         SaveWorkingMemory();